Not convinced here.
The problem is that Anthropic seems to be getting away with selling one thing and delivering another. You pay for Opus, you get something else etc.
Prompt was "read and update the config file with new data". This work on 4.6 takes <2 minutes to read the file, parse the new data, and patch.
Opus 5 Result: 43 minutes of pulling containers, running sandboxes, creating testing suites, which included evaluating the entire repo beyond the scope of the config file.
Both: one file modification
/on The prose is load-bearing unbearable — every sentence feels like it was engineered to sound profound rather than to be read.
I remember how I enjoyed agents between December and February, something started changing around March.
I thought models are getting dumber, but benchmarks were convincing opposite, initially I thought maybe they're quantizing models for day to day use, but Opus 4.8 and Opus 5 seems worse models than Opus 4.6
it really angers me every day
Opus 5 is a neverending chain of "Don't do that. Why did you do that? I've told you several times not to do that but you keep doing it."
"Thinking" for more than 10 minutes for every menial question.
And the prose it writes is horrendous, as if you're reading LinkedIn scammers. "The harsh truth! Two roads, one decision! Reality check!"
I keep finding myself typing “stop overcomplicating everything” multiple times a day as well.
Config file requires being updated with 3rd party data that's 2k of lines.
It's selective edits so instead of me reading the file, selecting the parts, validating syntax, documenting changes that would take 30 minutes considering the complexity i have AI do the work, review, validate, generate tickets, create pull requests and in 2 minutes I'm done.
Recently got approved at work for ChatGPT Pro so I could use Codex.
Blown away by the speed. It feels like using Claude Code for the first time again. I don't think Codex is doing anything revolutionary, just better handling of which requests should go to which model, and having faith in some of the "less powerful" models for more than you would think.
It seems the TUI coding experience is very much an open race. This is motivating me to look at other agents / harnesses as well (maybe Gemini, OpenCode, etc).
I've had WAY better results on medium effort.
IIUC, the consensus seems to be that anything more than medium effort is rarely worth it - and you far more often run into these extreme worst cases than you do with even the lowest effort levels. That definitely coincides with my anecdata.
It's really only worth it if you're hoping to win the lottery asking it to solve an Erdos problem.
For example I'm working on a project now where I need to keep in mind details from 3 separate source repos in different programming languages, along with probably a dozen important business-related documents that either corroborate the stuff in the source repos or add additional important context. It's a lot of details for even a human to manage, and when it comes to actually synthesizing plans and reports across this sprawling information environment, anything less than Opus 5 on Extra+ tends to miss important details and make bad recommendations or draw incorrect conclusions, which then poison subsequent context.
I suspect some kind of RAG-like memory system would greatly facilitate a project like this, but even with such a system I'm not confident that I could get away with an LLM that "thinks" less hard than this. I will say that slogging through the generated documents is kind of miserable and I have to repeatedly fork off side conversations to ask for clarification, but sometimes leads the model to "realize" it's made a mistake in all of its dense babbling, and it's all very hard to interpret. I never had much interest in trying GPT 5.6 until now.
If I have a user input and then sanitize and inject that into a prompt to do something, I have no idea how much that is going to cost at all and no real way to measure this properly. A parallel example is digital ocean or aws, i can go and measure/limit my compute/fs/memory/startup times/etc and while it can be impossible to get down to the last flop of money allocated - i can run things on a real budget with real constraints, opposed to an LLM where I have to .. prerun a sanitized user prompt through a tokenizer and then ask an LLM to guess what it may do and give token consumption estimates and then act on those in any sane manner for the user?
Perhaps i'm missing something to do realistic and static rails on things but I don't see a serious way at scale to use the token billing model handling things requiring a users free text input short of having to go pander to VC money to throw money at it until someone else figures it out.
*to clarify my rambling... We should be billed and given controls based on resource usage itself and not an opaque token concept on top of not being able to spin any knobs that control it's resource usage.
We sometimes test API serving configs in Claude Code before rolling them out, and one running now maps the numerical effort value differently.
That's why Claude may tell some of you it's at "10" on high. The scale isn't 0-100, the number isn't meaningful on its own, and the effort you selected is the effort you're getting. We've run in-depth evals to confirm this doesn't affect model performance.
This should be the same experience, but if you see a clear regression please hit /feedback and send me the ID. Will give credits.
"We sometimes test API serving configs in Claude Code before rolling them out, and one running now maps the numerical effort value differently. That's why Claude may tell some of you it's at "10" on high. The scale isn't 0-100, the number isn't meaningful on its own, and the effort you selected is the effort you're getting. We've run in-depth evals to confirm this doesn't affect model performance. This should be the same experience, but if you see a clear regression please hit /feedback and send me the ID. Will give credits."
LLM doesn't seem to be keen to put in effort either!
Is this AGI?
The chat-based models are obviously being lobotomized based on personal usage and general load (e.g. PST business hours are worst).
API doesn't seem to be affected by this.
I was approved.
3 months later, my approval was degraded into "in review" (revoked). I'm sure my account was flagged based on contents of debugging/researching firmwares/etc.
I opened a support ticket. No response. I opened another support ticket. No response.
1-2 weeks later, I got a response that I will not be re-approved and I need to reapply. No problem.
The page to reapply on does not allow me to re-apply because it my account is stuck in an "in review" status.
https://github.com/anthropics/claude-code/issues/84352
The community thinks it's a bug. I'm 95% sure it's not and a bunch of us who were previously approved had it revoked due to flagged content and will not be reapproved.
I switched to Codex + got TAC approved instantly and have not looked back. It's a shame. That's 100% separate from whatever the heck the quality of Opus 5's outputs are. The way it talks... insane. I would bet a good amount of money their next release will focus "reduced simplified responses" if I had to guess.
$2t company by the way
Meanwhile the Chinese models are "go ham dude"...
If it was not for capacity issues, Chinese models have a higher change to just dominate.
> $2t company by the way
It used to be that OpenAI and Anthropic had such a moat around them, that such a valuation was worth it. But these days, its gross overvalued (like so many).
The more stuff is being pulled like cyber verifications, downgrading effort levels, downgrading usage (OpenAI), the more people move to those Open Weight Chinese models.
A fun recent event ... https://opencode.ai/data/
When DeepSeek Flash 0731 came out and provided a massive jump in cheap inference capability. It resulted in a 10x increased OpenCode token usage.
It took a 2.5x to 5.0x price increase AND a reduction by 4x usage (later to 2x) usage, and several cheaper models + a free model, to push the traffic down.
Traffic towards open weight models is increasing, even if providers can not keep up with the influx of new customers. This is not something you want to see as two companies, trying to go for IPOs.
So the idea of stonewalling cyber capabilities, when the rest of the world is just doing whatever with open weight models, on their own hardware even! This entire strategy from Anthropic never made any sense.
In that case, the agent will respond incorrectly because it has no visibility into what reasoning mode it’s in.
"Opus 5" is actually Opus 4.8 in a trenchcoat, with new guardrails
"Opus 4.8" is the old Opus 4.6 with lipstick on it, with new guardrails
"Opus 4.6" which everybody used to love, is now actually running the old Opus 3.5...
How would we be able to tell?
The mood change in just six months is wild, in February this year Claude was the most liked LLM by far.
I think it's optimized around beating benchmarks and running fully autonomous in pursuit of a clearly defined goal. It makes sense: fan out aggressively, chase down every lead, but go depth first because that's easier for the LLM and you're either a sub-agent with a narrowly defined task or you're a top-level orchestrator agent with a /goal loop that will catch and fix errors and omissions on the second, third, fourth pass.
EDIT: my honest opinion; Anthropic is building a person, whereas everybody else (it seems) is building a tool.
I realized I was spending most of my tokens arguing with Opus and trying to get it to let go of stupid, lazy, obviously incorrect premonitions. I wound up canceling my 200, bought a pair of Sparks, and am running full fat DS4 Flash and so much happier. Done with being at the whim of these companies.
but the incentives of these companies to become profitable at any cost slipping into entire new types of dark patterns around token based billing seems gross
- charging for injected prompts and cot tokens
- changing default thinking effort to be higher
- training models to give longer winded answers that don’t say anything more of substance
- refusing to fulfill a request and still charging you
i wonder if you could ever just charge based of each user message and it so how breaks even across short and long replies
I have been using GLM-5.3 in my home setup and it is very good in comparison.
I didn't really care about their mobile app and the worst experiences were sometimes the sub-agents within OpenCode freezing and refusing to report their status (though this also happened with Kepler by the GitKraken folks).
I maintain a Claude subscription for Fable but seldom use it.
I'm building it so I'm totally open to feedback and would love to build something people really like.
The model providers are quite aligned with concerns like customer retention. These arguments only work if there is no competition. We exist in a marketplace of black boxes. There's not just "the one" you must suffer. You have options. You can build your own too.
They’re growing over 10x a year. They want users and revenue. In order to get users and revenue, they want to provide the smartest models at affordable prices. If they unnecessarily burn tokens, users will get less value and switch.
This thread is filled with competing comments about their monopolistic power and how when one model provider was no longer doing a good job people switched to a different one.
The competition in this market is ferocious!
Theoretically.
In reality, one sessions output tokens become the next sessions input tokens (at least if you continue the topic) so, its not as aligned as all that.
But the parent is right, when incentives are not aligned, friction will happen. Its inevitable.
https://code.claude.com/docs/en/admin-setup#set-up-usage-vis...
Goas back to factory towns, gift cards, game money or MtG.
Appreciate the outreach that you do! I love Claude, but I've been noticing reduced fidelity lately. Fable's likelihood of making a mistake increases or decreases based on the hour of the day and whether or not it's the weekend.
On a related note, and I'm happy to work on quantifying it, but qualitatively it feels like Fable's performance is noticeably poorer than initial release / launch.
I am wondering if this is the case because I use Claude via Claude Code to make a personalized care dashboard for my doctors to help me in managing my care.
I noticed in the upgraded filter announcement, https://www.anthropic.com/news/improving-fable-5-s-biology-s... ,
"In the case of Fable 5, when a classifier fires, the model re-routes the user’s request to Opus 5, a capable model that does not have the same level of biological capability as Fable 5 and which cannot provide as much assistance to a malicious user. This is the fallback that users see when their requests are blocked."
Is the end user informed every time their query is re-routed?
Or, can you confirm that there aren't scenarios where a user's outputs are degraded without telling them? As was the case for AI research during launch?
Why is it considered acceptable to test on paying customers without letting them know or giving them a way to opt out?
1. Self hosting
2. Chinese models
3. Running it locally. Requires upfront cost and compromises on TPS.
Well if you're not even going to put in the effort...
Wait. Let me do this for you.
/config model=boredumb effort=xhigh
/ultraplan "Build a frontier model that can run @1ktok/sec anywhere DOOM will run"Hell, I use 3 different providers, and I currently don't give a dime to Anthropic or OpenAI.
I see posts all the time on HN about which models from which providers offer the most bang-for-the-buck, and how to minimize token usage and still get optimal results, so it appears that competition is working.
- OpenAI
- Anthropic
- xAI (Grok)
- Moonshot (Kimi)
- Z (GLM)
- Meta (Muse)
- Deepseek
When you’re doing this 100K times per day you get an extremely good idea of what it costs. You also have all the tools to see when something starts changing quickly.
This change is for Claude Code the harness. If you’re using the API at scale and paying full price then you get exactly what you put into the request.
I am not a developer although I have written production code but that's not what I am talking about.
I know for a fact that I can do a task in a week. It may take 12 or 40 hours but it'll be done in a week. Now with AI? I can get that task done in a day. Anywhere from 3-12 hours.
You ship 3 hours in three with AI. The old number is meaningless now.
I agree that incentives are misaligned but there’s several competing model providers. If one gets funny with their costs people will jump ship, especially if the gap between the top 2 labs and everyone else keeps shrinking.
These safeguards already exist when they get a whiff that you might be using Claude to fix security issues. Doesn’t seem farfetched given the incentives I outlined that they would apply to this kind of abuse.
How loose those controls are becomes a market force.
I was a 4.6 acolyte from April til the fable drop, lost that quick, cancelled and took a break, came back a month later, tried opus 5 and liked it, so unpinned 4.6.
Results were great at first, and they're still not terrible, but I have noticed a regression in accuracy, so to speak, where I am pointing out issues that are quite obvious in review.
I pretty much use sonnet 5 low/medium when I have a plan to solve a simple problem and depending on scope, opus low/medium for more complex/bigger scope implementation, and only go high when it's very complex or I'm spitballing architecture/solutions and iterating plan. Never go xhigh or max.
The verbosity is insane though, opus 5 documents everything and just regurgitates whatever lead it to the design choice in there, which makes it more opaque because it's talking about something that was discussed once in a session that no one else can see (except their backend ofc)
I don't even try to steer it away from that with harness, because it doesn't work and just ends up agonizing over whether it should write some comment. Three paragraphs waffling on that on verbose output
I did however have it write a script that basically is git add -A -p for comments though, haha.
I'm $20/month, have all my telemetry toggles off, don't really over engineer prompt/context, just some basic skills for repeated patterns.
If I give it the specific instruction to "elide all revisions, corrections, and past mistakes" it usually works. You can also have Sonnet do a cleanup writing/style pass in a subagent. I impression is that Opus has been deliberately trained to keep track of all such revisions by default as a kind of ad-hoc memory mechanism. It's probably good for autonomous coding and beating benchmarks, and I presume reduces flailing when a separate session needs to pick up the work.
It was the wrong time for the GP to drop that subscription from $200 to $20, because $200 gets you a metric assload of cognition while $20 gets you nothing beyond what a local model running on your own graphics card can deliver.
Consistent, predictable behavior is valuable, even more so given the nondeterministic nature of LLMs. Nondeterminism combined with unpredictability might as well be randomness.
I'd guess you're deliberately exaggerating here, but still. I've never clocked the actual tokens/second, but I'm on the $20 plan and get ~15M tokens/month for fully utilized weekly quotas (checked couple months ago). Meanwhile the best I've been able to get locally was ~8 tokens/second with Qwen3.6 35B A3B, which is wildly painful for coding sessions and gets a maximum ~20M tokens in a month... if it's going 24/7.
Just wanted to stick some empirical data here, given that statement.
At $1500 that's 75 months of $20/mo Claude which are MUCH better models than you can run locally.
One challenge I run into is I run many agents at once, so local resource are a tiny fraction of the total inference we’re using. Every developer with his own Pro 20x, Max, etc. accounts is letting us hit literally trillions of tokens a month.
But just so we're both entirely clear on what an LLM is... it's a token prediction system.
it genuinely can't do things except recall things that have already existed.
People are having great success composing things together in new ways, but just like the english language has a finite number of sentences, and music has a finite number of chords: LLMs too are just combining things that have existed.
I don't want to sound condescending, it is remarkable how useful this technology is, but please don't evangelise them on capabilities that they genuinely can never have.
Laptop computers have incredible processing capabilities but nobody expects them to be able to walk your dog, no matter how useful they actually are at doing other things.
The point raised by this very article is that you can't depend on that. It's Flowers for Algernon As A Service.
I think we just need to decouple "doing better on benchmarks" and "actually more useful to me", since they've clearly diverged
There's a third perspective here: models are getting less useful, but overfitting to seeming useful to humans.
Imho, this is why analysis like TFA + third party cross-compatible harnesses (read: last mile UX) are so important to the leading labs optimizing for actual utility.
I'm suspicious enough of my subjective evaluation to believe a well-designed harness / verbiage could gaslight me into believing an objectively inferior model was superior. And at some point frontier labs are looking at the ROI of investing $1 in that vs actual model improvement.
The economics catching up with the providers in regards to how much compute they can burn per request and have it make sense for them financially?
A sort of model collapse where Opus 5 seems to love throwing out long paragraphs of text and it needs to be "fixed" by changing the output style and other patches.
I'm not sure, it might also catch up to Kimi K3 and GLM 5.3 and the models that I'm moving to from Anthropic.
>Opus 4.8 and Opus 5 seems worse models than Opus 4.6
After all we've heard about benchmark cheating, I'm earnestly not sure which or whether benchmarks are reliable anymore. But, beyond the models, I wonder if changes to their harnesses and/or instructions dumb them down. I have noticed models change their behavior, even when using the same version/effort. Sometimes for better. Sometimes for worse.
And, I have noticed a model go from really good to struggling. On 4.8 things were going well for a good stretch, so I did not switch to 5 when it came out. Even after hearing complaints about 5, 4.8 was still going well. Then, suddenly over the last few days, 4.8 seems to have nosedived. It feels similar now to the complaints I hear about 5.
In my case, it suddenly started ignoring my design guide, and introducing new fonts etc. It would even use several different fonts and sizes, as well as different margins for similar elements within the same page. It abandoned classes and started inlining styles. It started feeling random and, even after it realized it needed to go back to the design guide, it just continued with more of the same.
There seems to be something that happens after new model releases in both quality and behavior of previous models. It may not be immediately, but eventually there is frequently some regression.
In addition to that, while yes, 4.6 and 5.0 can solve problems, they do so differently. Sometimes 5.0 does better by a wide margin but that would be expected as they are supposed to be better.
If so, I would think that would result in worsened performance, not quality (unless you are also suggesting they may be quantizing).
models are incentivized by their makers to burn through as many tokens as they possibly can, so long as the customer doesn't cancel.
Opus models have degraded rapidly since March, with each release being considerably less useful and considerably more verbose. The language is no so weirdly florid it’s difficult to understand, and its logical conclusions are almost always suspect. It goes off on clearly bizarre snipe hunts to the point it feels like I’m using a gpt 3 model at times. It’ll announce that it’s about to embark on building something then just return control to the user and wait. You can also tell perceptibly when they’re reducing model quality to load shed - it becomes stupider and stupider to the point you’re better off dumping state and switching to codex or just turning in for the day and hoping they secured more capacity tomorrow.
It’s an absolute race to the bottom with Anthropic on virtually every level. I’ve rarely seen a company so rapidly accumulate good will in the developer community as they did around 4.6 in December and January. By March, it was inconceivable to use anything else. 4.8 was a bit of a wake up call to not put all your harness eggs in one basket. 5 is straight up time to cancel territory.
I actually manually set my model back to the older versions to get anything serious done. More and more I use codex for anything non trivial.
This isn’t about avarice by the provide trying to get more tokens and more utilization. They’ve over subscribed for capacity as it is. If they can produce better quality for less tokens they can charge a higher margin and will be paid if, which is a better economic strategy overall. This is something else. I suspect it’s actually the opposite, they’re finding ways to cut capacity demand in ways that leads to worse behavior that leads to more capacity demands, worse output, worse quality, and worse margins, worse, worse, worse.
Just as I never saw a company accumulate such positive developer good will so fast, I’ve never seen one squander it so fast too.
You may say sure but why? We could cook over an open fire too but we have microwaves and stoves and restaurants and protein shakes.
Every generation since fire to bronze to internal combustion engines has adopted the new technology, integrated it so deeply into their lives that we recreate by going camping, disconnecting, or playing with toys that resemble the past era of forgotten tech.
Knowing how hard something will be to do before attempting it is precisely the sort of impossible thing that it couldn’t do.
a) if something is possible
b) if something has been requested to take a long time ( a signal of abuse, like requesting illicit pictures in image generation)
is actually somewhat straightforward (I mean, if they are able to predict if something is a “substantial piece of work” as they seem to do already).
The mathematical proof thing is obviously marketing spin, you should pay more attention to what mathematicians are actually saying instead of hackernews folks.
These things are really good at being search engines and harnesses for iteration rather than some kind of advancing intelligence.
https://cacm.acm.org/research/formal-reasoning-meets-llms-to...